Skip to main content

Use Cases

The AI proxy's superpower is that you never touch vectors — you send raw text, images, or audio and it does the embedding for you. Every example below is the same recipe: raw input in, nearest matches out, with optional metadata filtering.

Your inputtext · imageaudio · a userAhnlichfind nearest vectors+ filter by metadataRelevantresultsrankedSame recipe powersSemantic searchRecommendationsCross-modal (text↔image)Clustering
Every use case below is the same three steps: turn your input into a vector, find the nearest stored vectors, and optionally filter by metadata. Only the data and the labels change.

1. Semantic Search with Raw Input​

A news website integrates Ahnlich AI with all-minilm-l6-v2. Instead of keyword search, users type:

SEARCH "climate change and food security" IN news_store WHERE (topic != sports)

  • The text query is embedded automatically.

  • Stored articles are embedded consistently.

  • Ahnlich DB returns top semantic matches, filtering irrelevant categories.

This provides conceptual search rather than exact word matching.

2. Cross-Modal Search (Text ↔ Image)​

A fashion platform configures:

  • Index Model = resnet-50 (for product images)

  • Query Model = all-minilm-l6-v2 (for user text queries)

When a user searches for a product:

GETSIMN 5 WITH [red summer dress] USING cosinesimilarity IN fashion_store

Ahnlich AI embeds the text query and compares it against image embeddings in the store. This allows retrieving visually similar dresses from the catalog — without the store owner needing to manually tag the images.

3. Personalized Recommendations​

Ahnlich AI can also transform user profiles or behaviors into embeddings automatically.

Example: an e-commerce platform using product_store:

GETSIMN 5 WITH [eco-friendly home products] USING cosinesimilarity IN product_store WHERE (status = in_stock)

Here:

  • The user query is embedded.

  • It’s matched against product embeddings.

  • The results are filtered using the predicate (status = in_stock).

This enables real-time, personalized product recommendations tailored to availability.

4. Multimodal Applications in Healthcare​

Ahnlich AI does not support mixing image and text embeddings in a single store. Each store is model-aware and tied to one input type (text, image, or audio).

For workflows such as CT scans and radiology reports, the recommended approach is to create two separate stores and link them using metadata fields like patient_id or report_id.

Pattern A Two Stores with Metadata Linking​

Click to expand
create an image store
CREATESTORE ct_image_store QUERYMODEL resnet-50 INDEXMODEL resnet-50 STOREORIGINAL;
CREATEPREDINDEX (patient_id, report_id) IN ct_image_store;

create a text report store
CREATESTORE report_store QUERYMODEL all-minilm-l6-v2 INDEXMODEL all-minilm-l6-v2 STOREORIGINAL;
CREATEPREDINDEX (patient_id, report_id) IN report_store;

insert CT scan
SET (([<image-vector>], {patient_id: "P123", report_id: "R789"})) IN ct_image_store;

insert report
SET ((["Findings: ground-glass opacities ..."], {patient_id: "P123", report_id: "R789"})) IN report_store;

query CT scans
GETSIMN 5 WITH [<image-vector>] USING cosinesimilarity IN ct_image_store;
fetch linked report
GETPRED (report_id = "R789") IN report_store;

Here, similarity search on ct_image_store finds related scans, and report_id links results to the associated report in report_store.

Pattern B — Cross-Modal Matching with CLIP​

Click to expand
-- create a cross-modal store (text ↔ image)
CREATESTORE clip_image_store QUERYMODEL clip-vit-b32-text INDEXMODEL clip-vit-b32-image STOREORIGINAL
CREATEPREDINDEX (patient_id, report_id) IN clip_image_store;

insert CT scan
SET (([<image-vector>], {patient_id: "P123", report_id: "R789"})) IN clip_image_store;

query with text (embedded via CLIP text model)
GETSIMN 5 WITH ["ground-glass opacity in left lung"] USING cosinesimilarity IN clip_image_store

5. Real-Time Assistance​

A chatbot connected to Ahnlich AI can provide real-time recommendations by retrieving similar past support tickets.

Example flow:

  1. The user submits a query: "I need help with my billing issue"

  2. The query is converted into embeddings automatically.

  3. Retrieve the top N similar past tickets using GETSIMN:

    GETSIMN 5 WITH ["I need help with my billing issue"] USING cosinesimilarity IN support_store

  4. Optionally, filter results using metadata with GETPRED:

    GETPRED (resolved = true) IN support_store

This workflow allows the chatbot to return relevant historical solutions with low latency, using similarity search combined with predicate filtering.